RU2731659C2 - Method and apparatus for training machine learning algorithm (mla) for creating content recommendations in a recommendation system and a method and apparatus for creating recommended content using a machine learning algorithm - Google Patents

Abstract

FIELD: computer equipment.

SUBSTANCE: invention relates to machine learning means for selection of recommended content for user of electronic device. Method is carried out by a recommendation server, which is accessible to an electronic device through a data network, recommendation server performs algorithm of machine learning after its training. Method includes: obtaining an indication of a plurality of user-element interactions, each interaction associated with a user and a digital element; based on a plurality of user-element interactions, creating a user-relevance assessment matrix; decomposition of the user-element relevance assessment matrix into the user matrix and the elements matrix, wherein the decomposition includes: initialisation of a matrix of elements using element vectors, vectors of elements have been created such that digital elements with identical content have identical vectors of elements, initialisation of user matrix using user vectors; iterative optimization of user matrix and elements matrix; maintaining an optimized matrix of elements.

EFFECT: technical result is widening the range of recommended content for a user.

17 cl, 6 dwg

Description

FIELD OF TECHNOLOGY

[001] The present technology relates to recommendation systems in general, and more specifically to a method and apparatus for training a machine learning algorithm (mla) for generating content recommendations in a recommendation system and a method and apparatus for generating recommended content using a machine learning algorithm.

LEVEL OF TECHNOLOGY

[002] Various global or local communication networks (Internet, World Wide Web, local area networks and the like) offer the user a large amount of information. Information includes contextual sections such as, but not limited to, news and current events, maps, company information, financial information and resources, traffic information, games, and entertainment information. Users use a variety of client devices (desktop, laptop, laptop, smartphone, tablets, and the like) to access rich information content (for example, images, audio and video files, animated images and other multimedia content of similar networks).

[003] The amount of information available on various Internet resources has grown exponentially over the past few years. Various solutions have been developed that allow the common user to find the information he / she is looking for. An example of such a solution is a search engine. Examples of search engines include search engines such as GOOGLE ™, YANDEX ™, YAHOO! ™ and others. The user can access the search engine interface and confirm the search query related to the information the user wants to find on the Internet. In response to a search query, search engines provide a ranked list of search results. A ranked list of search results is generated based on various ranking algorithms that are implemented in a particular search engine and that are used by the user searching. The overall goal of such ranking algorithms is to present the most relevant results at the top of the ranked list, and the less relevant results at the lower positions in the ranked list of search results (and the least relevant search results will appear at the bottom of the ranked list of search results).

[004] Search engines are usually a good search tool when the user knows in advance what he (s) wants to find. In other words, if the user is interested in getting information about the most popular places in Italy (ie the search topic is known), the user can enter a search query: "Most popular places in Spain". The search engine will provide a ranked list of Internet resources that are potentially relevant to the search query. The user can then browse the ranked list of search results in order to obtain information in which he is interested, in this case, about the places visited in Spain. If the user is not satisfied with the presented results for any reason, the user can perform a secondary search by specifying the query, for example, "the most popular places in Spain in summer", "the most popular places in the south of Spain", "The most popular places in Spain for a romantic getaway" ...

[005] There is another approach in which the user is allowed to discover the content and, more specifically, is allowed to display and / or recommend content that the user was not explicitly interested in searching for. In a sense, such systems recommend content to the user without a separate search query, based on the user's explicit or implicit interests.

[006] Examples of such systems are the FLIPBOARD ™ recommendation system, which aggregates and recommends content from various social networks. The FLIPBOARD recommendation system provides content in a "journal format" where the user can "flip" through pages of recommended / aggregated content. Recommender systems collect content from social media and other websites, present it in a magazine format, and allow users to “flip through” social news feeds and web site feeds that partner with a company to effectively “recommend” the content to the user. even if the user has not explicitly expressed their interest in the particular content.

[007] Another example of a recommendation system is the Yandex.Zen ™ system. The Yandex.Zen ™ recommendation system recommends digital content (for example, articles, news, and videos in a personalized feed on the Yandex Browser home screen). When a user views content recommended by the Yandex.Zen server, the server receives explicit (by asking the user if he (s) wants to see more similar content in his feed) or implicit (by observing the user's interaction with the content) feedback. With the help of user feedback, the Yandex.Zen server constantly improves the content recommendations that are provided to this user.

[008] To generate ranked search results in a search engine or a list of recommended resources in a conventional recommendation system, the respective systems use machine learning algorithms to select search results and / or content recommendations. There are various machine learning algorithms known in the art and used in search engines and recommendation engines. As you know, conventional machine learning algorithms are first "trained" using a training set (which is flagged or unlabeled) to create a machine learning algorithm formula, which is then used to determine at runtime the output based on the system input.

[009] In a typical recommendation system, the recommendation system machine learning algorithm selects a set of potential recommended items from a set of potential resources for recommended items (for a specific user). Typically, a machine learning algorithm that is associated with such a recommendation system acquires (over time) knowledge of the various potential content item resources, the user's previous interactions with the various potential content item resources (as well as generated recommendations), and uses this knowledge to create recommendations for a specific user of the recommendation system.

[0010] One of the algorithms that are used to generate content recommendations is collaborative filtering. Collaborative filtering is based on relevance scores, which are calculated based on previous interactions of many users with many digital items. These interactions can be associated with explicit or implicit user feedback that is used to generate a given relevance score for a given user interaction with a given digital item.

[0011] To create recommendation systems, a relevance score dataset that is associated with various potentially recommended numeric items is used as a matrix for processing. This matrix is usually small (i.e. not every digital item has been interacted with by enough users, and not every user has interacted with every potentially recommended digital item). The usual traditional approach involves matrix decomposition to fill in missing relevance scores.

[0012] In linear algebra, matrix decomposition can be performed using Singular Value Decomposition (or SVD for short). In accordance with the SVD algorithm, any matrix can be represented as an expansion according to the formula:

[0013] As a result of the singular value decomposition algorithm, a conventional matrix of relevance scores (which are based on user-digital item interactions) can be divided into a user matrix and a digital item matrix.

[0014] When creating a recommendation system, an approximate matrix decomposition is usually used. There are two common approaches to performing approximate matrix decomposition: gradient descent and alternating least squares (ALS). Therefore, the ALS approach allows for both explicit and implicit feedback for recommendation systems.

[0015] Broadly speaking, the ALS approach can be represented as follows. The relevance score matrix ® is presented as R≈UV + T. Following the usual ALS approach, the values of the initial matrices U and V are randomly selected and filled with zeros. The ALS approach further consists of two stages: stage U fixes V and tunes U, stage V fixes U and tunes V. After each iteration, the decomposition of U and V approaches R.

[0016] US Patent Application 2015/0052003 describes a method for providing personalized content recommendations using scalable random matrix decomposition. Other embodiments of related systems and methods are also described.

[0017] Patent Application US 8,983,888 describes a method for efficiently decomposing a matrix in a recommendation system. Usage data for a large set of users pertaining to a set of elements is provided in the R matrix. To reduce computational power, the usage matrix is sampled to represent the reduced matrix R '. R 'is decomposed into a user matrix U' and a V element matrix. The user vectors in U 'and V are initialized and then iterated over to arrive at an optimal solution. The reduced matrix can be decomposed using computational resources, for example, one computer device. Therefore, the full user matrix U is formed by fixing V and analytically minimizing the error in the equation UV = R + error. The computations of this analytical solution can be distributed among a set of computing devices, for example, using the map-and-reduce technique (Google's framework used for parallel computing on very large datasets in computer clusters). Each computing device solves the equation for different respective subsets of users.

[0018] Patent application US 8,676,736 describes a system and method for providing personalized featured items in a communication system that includes a server and a plurality of client devices. On the server, a plurality of user rating vectors are obtained from a plurality of client devices and are aggregated into a rating matrix, which will then be decomposed into a user characteristics matrix and an item characteristics matrix, and the resulting user characteristics and item characteristics matrices approximate the user rating matrix. The decomposition includes the steps of the ALS1 or IALS2 algorithms, including: initializing the user characteristics matrix and the element characteristics matrix with predetermined initial values; alternative optimization of the user characteristics matrix and the element characteristics matrix until the final state is reached. The element characteristics matrix is transmitted from the server to at least one client device, and the predicted action vector is created as a product of the corresponding user characteristics vector and the element characteristics matrix. At least one element is selected for recommendation to the user among the elements associated with the predicted rating vector.

[0019] An article entitled “Fast ALS-based Matrix Factorization for Explicit and Implicit Frrdback Datasets” by Pilgie et al. The RecSys conference September 26-30, 2010 (Barcelona, Spain) describes the ALS algorithm as an decomposition algorithm for explicit and implicit feedback based on recommendation systems.

DISCOVERY TECHNOLOGY

[0020] the Objective of the proposed technology is to eliminate at least some of the disadvantages inherent in the prior art. Embodiments of the present technology can provide and / or expand the boundaries of the present approaches and / or methods of achieving the goals and objectives of the present technology.

[0021] Embodiments of the present technology were developed on the assumption of the inventors that machine learning algorithms (or special algorithms used to train machine learning algorithms) used to select content recommendations can consume a lot of computer computing resources. Without being limited to any particular theory, the developers of the present technology believe that randomly selecting the initial matrices for the alternation least squares (ALS) approximation results in the overall process is a time-consuming and computationally intensive operation. The same applies to the initial variant with zeros for the initial matrices in the ALS fit.

[0022] Broadly, embodiments of the present technology aim to optimize the ALS approximation of the matrix decomposition for the recommendation system. More specifically, embodiments of the present technology have been developed based on the developer's judgment that when the ALS algorithm is used, the original element matrix can be populated with element vectors that at least partially represent actual digital elements and are at least approximation to a set of latent factors of digital elements. The technical effect of at least some embodiments of the present technology is that the least squares interleaving algorithm requires fewer iterations for approximate decomposition. In some embodiments of the present technology, the technical effect may be that the approximate decomposition takes less time to complete. In some other embodiments of the present technology, the technical effect may be that less processing power is required to perform the approximate decomposition. In other words, the technical result is to optimize the learning process of the machine learning algorithm by accelerating the matrix decomposition and reducing the required computing power.

[0023] In a broad sense, a server (eg, a recommendation system server) first obtains a history of user interactions with multiple digital items. In some embodiments of the present technology, the plurality of digital elements may be textual, such as articles, books, other texts, and so on. Logs provide indications of user interactions with multiple digital elements (or, more accurately, an indication that some users interacted with relevant digital elements, as well as the nature of their interactions). In some embodiments of the present technology, the interaction may include explicit user feedback regarding a digital item (eg, "like"); in other embodiments of the present technology, the interaction may not include explicit user feedback. However, user interaction can contain implicit user feedback (for example, the amount of time the user spent interacting with the digital content, the number of “reposts” of the digital item by the user, the number of comments the user left in relation to the digital item, and so on. ).

[0024] The server further converts the indication of user interactions into a user-item relevance score matrix. User-to-item relevance scores can be generated based on an indication of multiple users and a specific digital item. The type of user interaction is not limited in any way and can be (without setting restrictions): time spent interacting with a digital element, uploading a digital element, reposting a digital element, adding a digital element to bookmarks, leaving a comment related to a digital element, leaving a "like" »The digital element and updating the digital element. Based on the indication of user interactions, the server determines how relevant a given digital source is for a given user (this user has this user profile).

[0025] The server then performs decomposition of the user-item relevance score matrix. More specifically, the server can use a singular value decomposition (SVD) algorithm to split the user-to-item relevance score matrix into a user-item matrix and an item-matrix. Embodiments of the present technology perform an ALS algorithm to decompose a matrix of user-item relevance scores. In accordance with embodiments of the present technology, during the initialization of the primary matrices for the ALS algorithm, the initial matrix of elements is populated with element vectors that represent the actual digital elements. In those embodiments of the technology where the digital elements are textual, the element vectors are text of the textual digital elements.

[0026] The element vectors of text numeric elements can be created using word embedding techniques. The type of word embedding is not limited in any way, and some examples may include, but are not limited to, word2vec algorithms or Latent Dirichlet Allocation (LDA). In general, the selected one or more word embedding algorithms are configured to create a vector of elements in such a way that the vectors of elements of digital elements that have similar texts are essentially similar. In other words, the selected one or more word embedding algorithms provide matching of similar text numeric elements in similar element vectors (i.e., element vectors that are close to each other in the multidimensional space that is used to match digital documents and element vectors).

[0027] The initialization of the user matrix with user vectors can be performed in the usual manner, i. E. using random values for the initial user vectors, using zeros as the user initial vectors, and so on.

[0028] the Server then iterates over the user matrix and the element matrix. After optimization, the server saves the optimized matrix of elements.

[0029] One object of the present technology is a method for teaching a machine learning algorithm to select recommended content for a user of an electronic device, the method is performed on a recommendation server accessible to an electronic device via the Internet, the recommendation server executes the machine learning algorithm after learning. The method includes: obtaining an indication of a plurality of user-item interactions, each interaction being associated with a user and a digital item; based on a variety of user-element interactions, creating a matrix of user-element relevance scores; decomposition of the matrix of user-element relevance scores into a matrix of users and a matrix of elements, and the decomposition includes: initialization of the matrix of elements using vectors of elements, the vectors of elements were created in such a way that digital elements with similar content have similar vectors of elements, initialization of the matrix of users with using user vectors; iterative optimization of the user matrix and the matrix of elements; saving an optimized matrix of elements.

[0030] In some embodiments of the method, the method further includes: upon receiving a request for recommended content from an electronic device, obtaining a user profile associated with the electronic device; and selecting at least one recommended content item based on the user profile and the optimized item matrix.

[0031] In some embodiments of the method, the selection includes reconstructing a user-member rating matrix with an optimized member matrix and a user profile to create a reconstructed user-member matrix.

[0032] In some embodiments of the method, each user-item pair of the reconstructed user-item matrix is associated with a corresponding user-item relevance score, the corresponding user-item relevance score represents the digital item's relevance to the user.

[0033] In some embodiments of the method, the user profile is a vector based on the user's browsing history.

[0034] In some embodiments of the method, initializing the user matrix with user vectors includes populating the user matrix with random values of the user vectors.

[0035] In some embodiments of the method, initializing the user matrix with user vectors includes populating the user matrix with zero user vector values.

[0036] In some embodiments of the method, after iterative optimization, the user matrix is omitted.

[0037] In some embodiments of the method, the digital item is a text numeric item and wherein the item vectors have been created using the word embedding method.

[0038] In some embodiments of the method, the method further includes creating vectors of elements using a word embedding method, the word embedding method is at least one of word2vec or latent Dirichlet placement (LDA).

[0039] In some embodiments of the method, the decomposition is performed using SVD singular decomposition analysis).

[0040] In some embodiments of the method, the decomposition further includes decomposition analysis, the decomposition analysis is performed using an alternating least squares (ALS) algorithm.

[0041] In some embodiments of the method, obtaining an indication of a plurality of user-item interactions includes obtaining an indication of a plurality of user-item interactions from user browsing histories.

[0042] In some embodiments of the method, a user-item interaction is: time spent interacting with a digital item, downloading a digital item, reposting a digital item, adding a digital item to bookmarks, leaving a comment associated with a digital item, leaving a "like" »The digital element and updating the digital element.

[0043] In some embodiments of the method, iterative optimization of the user matrix and the matrix of elements is performed until a predetermined value of a predetermined value is reached.

[0044] In some embodiments of the method, the predetermined value is one of: Root mean square deviation (RMSE) and mean absolute error (MAR).

[0045] Another aspect of the present technical solution is a server. The server includes: a data carrier: a network interface capable of communicating via a data transmission network; a processor functionally connected to a data carrier and a network interface, the processor is configured to perform: receiving an indication of a plurality of user-element interactions, each interaction is associated with a user and a digital element; based on a variety of user-element interactions, creating a matrix of user-element relevance scores; decomposition of the matrix of user-element relevance scores into a matrix of users and a matrix of elements, and the decomposition includes: initialization of the matrix of elements using vectors of elements, the vectors of elements were created in such a way that digital elements with similar content have similar vectors of elements, initialization of the matrix of users with using user vectors; iterative optimization of the user matrix and the matrix of elements; storing the optimized matrix of elements in the data carrier. As used herein, "server" refers to a computer program running on appropriate hardware that is capable of receiving requests (eg, from client devices) over the network and executing those requests or initiating the execution of those requests. The hardware can be one physical computer or one physical computer system, but neither is required for this technology. In the context of this technology, the use of the expression "server" does not mean that every task (for example, received commands or requests) or any specific task will be received, executed or initiated for execution by the same server (that is, by the same software software and / or hardware); this means that any number of pieces of software or hardware devices can be involved in receiving / transmitting, executing or initiating the execution of any request or the consequences of any request related to a client device, and all of this software and hardware can be one server or multiple servers both are included in the expression "at least one server".

[0046] In the context of the present description, "client device" refers to a hardware device capable of operating with software suitable for solving the corresponding task. Thus, examples of client devices (among others) include personal computers (desktops, laptops, netbooks, etc.) smartphones, tablets, and network equipment such as routers, switches and gateways. It will be appreciated that a device that behaves like a client device in the present context can behave like a server with respect to other client devices. The use of the term "client device" does not exclude the possibility of using multiple client devices to receive / send, perform or initiate the execution of any task or request, or the consequences of any task or request, or steps of any of the above method.

[0047] In the context of the present description, "database" means any structured set of data that is independent of a particular structure, database management software, computer hardware on which the data is stored, used or otherwise made available for use. The database may reside on the same hardware running the process that stores or uses the information stored in the database, or the database may reside on separate hardware, such as a dedicated server or multiple servers.

[0048] In the context of the present description, "information" includes information, any information that may be stored in a database. Thus, information includes, among others, audiovisual works (images, videos, sound recordings, presentations, etc.), data (location data, digital data, etc.), text (opinions, comments, questions , messages, etc.), documents, tables, word lists, etc.

[0049] In the context of the present description, "component" means software (corresponding to a particular hardware context) that is necessary and sufficient to perform the specified specified function (s).

[0050] In the context of the present description, "computer-usable computer storage media" means absolutely any type and nature of media, including RAM, ROM, disks (CDs, DVDs, floppy disks, hard drives, etc.), USB flash drives, solid state drives, magnetic tape drives, etc.

[0051] In the context of the present description, the words "first", "second", "third", etc. are used as adjectives solely to distinguish the nouns to which they refer from each other, and not for the purpose of describing any specific relationship between these nouns. So, for example, it should be borne in mind that the use of the terms "first server" and "third server" does not imply any ordering, assignment to a particular type, history, hierarchy or ranking (for example) of servers / between servers, as well as their use (in itself) does not imply that a "second server" must necessarily exist in a given situation. Hereinafter, as indicated here in other contexts, the reference to the "first" element and the "second" element does not exclude the possibility that they are the same actual real element. So, for example, in some cases, the "first" server and the "second" server may be the same software and / or hardware, and in other cases, they may be different software and / or hardware.

[0052] Each embodiment of the present technology pursues at least one of the aforementioned purposes and / or objects, but all are optional. It should be borne in mind that some of the objects of this technology, obtained as a result of attempts to achieve the above-mentioned goal, may not meet this goal and / or may meet other purposes, not separately specified here.

[0053] Additional and / or alternative characteristics, aspects and advantages of embodiments of the present technical solution will become apparent from the following description, the accompanying drawings and the accompanying claims.

BRIEF DESCRIPTION OF DRAWINGS

[0054] For a better understanding of the present technology, as well as its other aspects and features, reference is made to the following description, which is to be used in conjunction with the accompanying drawings, where:

[0055] FIG. 1 illustrates a system suitable for implementing non-limiting embodiments of the present technology.

[0056] FIG. 2 is a screen shot of a recommendation interface made in accordance with a non-limiting embodiment of the present technology, the recommendation interface is shown on the screen of an electronic device of the system shown in FIG. 1, the electronic device is in the form of a smartphone.

[0057] FIG. 3 illustrates a plurality of user-item interactions (each user-item interaction is associated with a user and a digital item, a plurality of user-item interactions are handled by the system of FIG. 1 in accordance with some non-limiting embodiments of the present technology.

[0058] FIG. 4 is a schematic diagram of a user-item relevance rating matrix, a user-item relevance rating matrix generated by the system server of FIG. 1 through a variety of user-item interactions similar to those shown in FIG. 3, the user-item relevance score matrix is implemented in accordance with non-limiting embodiments of the present technology.

[0059] FIG. 5 is a schematic diagram of the process of decomposition of the user-element relevance assessment matrix into the users matrix and the element matrix, the process is performed by the server of the system shown in FIG. 1, in accordance with some non-limiting embodiments of the present technology.

[0060] FIG. 6 is a flowchart of a method for training a machine learning algorithm, the method being performed by the server shown in FIG. 1, in accordance with some non-limiting embodiments of the present technology.

IMPLEMENTATION

[0061] FIG. 1 is a schematic diagram of a system 100 implemented in accordance with non-limiting embodiments of the present technology. It is important to keep in mind that the following description of system 100 is a description of exemplary embodiments of the present technology. Thus, all of the following description is presented only as a description of an illustrative example of the present technology. This description is not intended to define the scope or delineate of this technology. Some useful examples of modifications to system 100 may also be encompassed by the following description. The purpose of this is also solely to help understand, and not to define the scope and boundaries of this technology. These modifications are not intended to be an exhaustive list, and those skilled in the art will appreciate that other modifications are possible. In addition, it should not be interpreted to mean that where it has not already been done, i.e. where no modification examples have been presented, no modification is possible, and / or that what is described is the only embodiment of this element of the present technical solution. As one skilled in the art will appreciate, this is most likely not the case. In addition, it should be borne in mind that the system 100 is, in some specific manifestations, a fairly simple embodiment of the present technology, and in such cases, this embodiment is presented here for the purpose of ease of understanding. As will be understood by one skilled in the art, many embodiments of the present technology will be much more complex.

[0062] In general, the system 100 is configured to provide content recommendations to the user 102 of the system 100. The user 102 may be a subscriber to the recommendation service that the system 100 provides. However, the subscription does not need to be billed or explicit. For example, user 102 can become a subscriber by downloading a recommendation from system 100, by registering and providing a username / password combination, by registering and providing user preferences, and so on. Therefore, any variant of the system capable of creating content recommendations for a given user can be adapted by a specialist to carry out embodiments of the present technology after a specialist has read the present description. In addition, system 100 can be described with an example of system 100, which is a recommendation system (hence, system 100 may be referred to as “recommendation system 100” or “prediction system 100” or “training system 100”). However, embodiments of the present technology may also be applied to other types of systems 100, as will be described in more detail below.

[0063] System 100 includes electronic device 104, electronic device 104 is associated with user 102. Thus, electronic device 104 may sometimes be referred to as a "client device", "end-user device" or "client electronic device". It should be noted that the fact that the electronic device 104 is associated with the user 102 does not imply any particular mode of operation, as well as the need to log in, be registered, or the like.

[0064] Embodiments of the electronic device 104 are not particularly limited, but personal computers (desktop computers, laptops, netbooks, etc.), wireless communication devices (smartphones, mobile phones, tablets, etc.) can be used as an example of the electronic device 104. .), as well as network equipment (routers, switches or gateways). Electronic device 104 comprises hardware and / or application software and / or system software (or a combination thereof) as is known in the art to execute recommender 106. In general, the task of recommender 106 is to allow a user to obtain (or what -Otherwise have access) to the content recommendations that are provided by the system 100, as will be described in more detail below.

[0065] The implementation of the recommendation application 106 is not particularly limited. One example of an implementation of the recommendation application 106 is a user accessing a website conforming to the recommendation system to access the recommendation application 106. For example, the recommendation application 106 may be invoked by entering (or copy-pasting or selecting a link) the URL associated with service of recommendations. Alternatively, the recommendation application 106 may be an application downloaded from a so-called app store, such as APPSTORE ™ or GOOGLEPLAY ™, and installed / used on the electronic device 104. It is important to keep in mind that the recommendation application 106 can be invoked by any other means ... In other additional embodiments of the technology, the functions of the recommendation application 106 may be embedded in another application, such as a browser application (not shown), and so on. For example, recommendation application 106 may execute as part of a browser application, for example, when user 102 first launches a browser application, the functionality of recommendation application 106 may be executed.

[0066] In general, recommendation application 106 includes a recommendation interface 108, the recommendation interface 108 being displayed on a screen (not separately numbered) of the electronic device 104. FIG. 2 is a screen shot of a recommendation interface 108 implemented in accordance with a non-limiting embodiment of the present technology (an example of a recommendation interface 108 is shown on the screen of an electronic device 104, which is implemented as a smartphone).

[0067] In some embodiments of the present technology, the recommendation interface 108 is displayed when the user 102 of the electronic device 104 activates (i.e., launches, uses, runs in the background, and so on) the recommendation application 106. Alternatively, the recommendation interface 108 may be presented when the user 102 opens a new browser window and / or activates a new tab in the browser application. For example, in some embodiments of the present technology, the recommendation interface 108 may activate a "home screen" in a browser application.

[0068] Suggestion interface 108 includes search interface 202. Search interface 202 includes search query interface 204. The search query interface 204 can be implemented as an "omnibox" that allows you to enter a search query to conduct a search, or a network address (eg, a Uniform Resource Locator (URL)) to identify a resource (eg, a website) to be invoked. However, the search query interface 204 may be implemented to obtain one or both of: a search query record to perform a search, or a network address (e.g., a Uniform Resource Locator (URL)) to identify a resource (e.g., a website) that will summoned.

[0069] The advisory interface 108 further includes a link interface 206. The link interface 206 includes a plurality of fragments 208 - eight of which are shown in FIG. 2 - only two are numbered in FIG. 2 - the first fragment 210 and the second fragment 212.

[0070] Using, for example, the first chunk 210 and the second chunk 212 - each of the plurality of chunks 208 includes (or acts as) a link to (i) a website marked as "favorites" or otherwise marked by user 102 , (ii) a previously visited website, or (iii) the like. The plurality of fragments 208 in the illustrated embodiment are visually presented to the user 102 in the form of square buttons with the logo and / or the name of the presented resource, the logo and the name allow the user 102 to identify to which resource each of the plurality of fragments leads (not separately numbered). However, it is important to keep in mind that the visual representation of some or all of the plurality of fragments 208 may be different. Thus, some or all of the plurality of fragments 208 may be implemented as buttons of other shapes, such as a list of hyperlinks, and so on.

[0071] For example, the first snippet 210 includes a link to the TRAVELZOO ™ website, and the second chunk 212 includes a link to the personal livejournal website. Needless to say, the number and content of individual chunks from the plurality of chunks 208 is not specifically limited.

[0072] For example, the number of tiles in the plurality of tiles 208 may be preselected by the vendor of the recommendation application 106. In some embodiments of the present technology, the number of tiles in the plurality of tiles 208 is preselected based on the size and / or resolution of the screen of the electronic device 104 that performs the recommendation application 106. For example, a first number of tiles may be preselected for an electronic device 104 that is implemented as a smartphone, a second number of tiles may be preselected for an electronic device 104 that is implemented as a tablet, and a third number of tiles may be preselected for an electronic device 104, which is implemented as a laptop or desktop computer.

[0073] The recommendation interface 108 further includes a set of recommended content 214. Recommended content set 214 includes one or more content recommendation elements, for example, first content recommendation element 216 and second content recommendation element 218 (second content recommendation element 218 is only partially visible in FIG. 2). Naturally, the content recommendation set 214 may have a large number of content recommendation items. Within the embodiment shown in FIG. 2, and those technology embodiments where more than one recommended content item is present, the user 102 can scroll through the recommended content set 214. Scrolling can be done in any suitable way. For example, the user 102 can scroll through the contents of the recommended content set 214 by activating a mouse (not shown), a keyboard key (not shown), or interacting with a touch screen (not shown) of the corresponding electronic device 104.

[0074] The example shown in FIG. 2 is one possible implementation of the recommendation interface 108. Another embodiment of the recommendation interface 108, as well as a description of how the user 102 can interact with the recommendation interface 108, is provided in a co-owned Russian patent application entitled "PC-USED METHOD FOR CREATING A RECOMMENDED CONTENT INTERFACE "filed on May 12, 2016 under the number 2016118519; the contents of which are incorporated herein in full by reference.

[0075] How exactly the recommended content set 214 is created will be described in more detail below.

[0076] Returning to the description of FIG. 1, electronic device 104 is operatively coupled to a communications network to access a recommendation server 112. In some non-limiting embodiments of the present technology, communications network 110 may be the Internet. In other embodiments of the present technical solution, the communication network 110 may be implemented differently as a wide area data network, a local data network, a private data network, and the like. The data link (not separately numbered) between the electronic device 104 and the data network 110 is implemented in such a way that it will depend, among other things, on how the electronic device 104 is implemented. By way of example, and not limitation, in these embodiments of the present technology, in cases where the electronic device 104 is a wireless communication device (such as a smartphone), the data line is a wireless data network (for example, but not limited to, a 3G data line, a 4G data line, Wireless Fidelity, or shortly WiFi®, Bluetooth®, etc.). In those examples where the electronic device 104 is a laptop, the communication link can be either wireless (Wireless Fidelity or WiFi® for short, Bluetooth®, etc.) or wired (Ethernet-based connection).

[0077] The recommendation server 112 may be a conventional computer server. In an exemplary embodiment of the present technical solution, the advisory server 112 may be a Dell ™ PowerEdge ™ server running a Microsoft ™ Windows Server ™ operating system. Needless to say, the recommendation server 112 can be any other suitable hardware, application software, and / or system software, or a combination thereof. In the illustrated, non-limiting embodiment of the present technology, the recommendation server 112 is a single server. In other non-limiting embodiments of the present technical solution, the functionality of the image recommender server 112 may be split and performed by multiple servers.

[0078] The recommendation server 112 includes a processing module 114 The processing module 114 is connected to or otherwise accesses the recommended content selection module 116. Recommended content selection module 116 has access to storage device 118. The processing unit 114 is further associated with the content detection unit 115. The operation of the recommendation server 112 and its components will be described in more detail below.

[0079] A plurality of network resources are also connected to the UE network, including a first network resource 124, a second network resource 126, and a plurality of additional network resources 128. The first network resource 124, a second network resource 126, and a plurality of additional network resources 128 are network resources. accessible to electronic device 104 (as well as other electronic devices potentially present in system 100) via a communications software network. The respective contents of the first network resource 124, the second network resource 126, and the plurality of additional network resources 128 are not particularly limited.

[0080] A given first network resource 124, a second network resource 126, and a plurality of additional network resources 128 may contain (or, in other words, host) digital content (i.e., one or more digital items from one or more digital items having one or several types of digital content). In some embodiments of the present technology, the content of the digital element may include, but is not limited to: audio content for streaming or download, video content for streaming or download, news, blocks, information about various government institutions, information about interesting places, thematically clustered content (for example, content relevant to those interested in kickboxing), other multimedia content, and so on.

[0081] In other embodiments of the present technology, the content of the digital elements located on the first network resource 124, the second network resource 126, and a plurality of additional network resources 128 is textual. Examples of text content elements include, but are not limited to: news, articles, blogs, information about various government institutions, information about points of interest, thematically clustered content (for example, content that is relevant to those interested in kickboxing), and so on. It should be noted that the term "textual" content does not mean that a given digital element includes only text and excludes other types of multimedia elements. In contrast, a given textual digital element includes textual elements as well as potentially another type of multimedia elements. For example, a given textual digital content element that is an article can contain text and photographs. As another example, a textual digital content item that is a blog can contain text and embedded video elements.

[0082] Content is potentially "available" to electronic device 104 in various ways. For example, a user 102 of an electronic device 104 may use a browser application (not shown) and enter a Uniform Resource Locator (URL) associated with one of the first network resource 124, the second network resource 126, and a plurality of additional network resources 128. Alternatively, the user 102 of the electronic device 104 may search with a search engine (not shown) to explore the contents of one or more of the first network resource 124, the second network resource 126, and a plurality of additional network resources 128. As previously mentioned, this is convenient if the user 102 knows in advance what content the user 102 is interested in.

[0083] In accordance with embodiments of the present technology, the recommender application 106 may recommend content items available from one of the first network resource 124, the second network resource 126, and a plurality of additional network resources 128 to the user 102, the content items about which the user 102 can do not know beforehand. The recommendation server 112 is configured to perform content selection for one or more of the recommended items that will be presented to the user 102 through the recommendation application 106. More specifically, the processing module 114 is configured to (i) receive a request from the electronic device 104 for content recommendations 150, and (ii ) in response to a request, generate a content recommendation message 152 specifically configured for a user 102 associated with an electronic device 104. Processing module 114 may further coordinate the execution of various procedures described herein as performed by content discovery module 115 and recommended content selection module 116, eg.

[0084] In some embodiments of the present technology, a request for content recommendations 150 may be generated in response to the user 102 providing an explicit indication of his / her desire to receive content recommendations. For example, the recommendation interface 108 may provide a button (or other activatable element) that allows the user 102 to express their desire for new or updated content recommendations. As a non-limiting example, the recommendation interface 108 may provide an actionable button that says “request content recommendations”. Within these embodiments of the technology, a request for content recommendations 150 may be considered an “explicit request” in the sense that the user 102 explicitly expresses a request for content recommendations.

[0085] In other embodiments of the present technology, a request for content recommendations 150 may be generated in response to an explicit indication from user 102 that the user wants to receive content recommendations. In some embodiments of the present technology, a request for content recommendation 150 may be generated in response to the user 102 launching the recommendation application 106.

[0086] Alternatively, in those embodiments of the present technology where the recommendation application 106 is implemented as a browser (eg, GOOGLE ™ browser, YANDEX ™ browser, a YAHOO! ™ browser, or any other proprietary or commercially available browser application), a recommendation request 150, the content may be created in response to the user 102 opening a browser application and may be created, for example, without the user 102 taking any action other than activating the browser application. As another example, a request for content recommendations 150 may be generated in response to a user 102 opening a new tab in an already open browser application, and may be created, for example, without the user 102 taking any action other than activating a new browser tab. In other words, a request for content recommendations 150 may be generated even without the user 102 knowing that he may be interested in receiving a content recommendation.

[0087] As another example, a content recommendation request 150 may be created in response to a user 102 selection of a specific browser application element, and may be created, for example, without the user 102 taking any action other than activating / selecting a specific browser element. applications.

[0088] Examples of a specific browser application element include, without limitation:

• Address bar of the browser application string

• The search bar of the browser application and / or the search bar of the search engine website accessible in the browser application

• Omnibox (related address and search bar of the browser application)

• Panel of favorites or recently visited network resources

• Any other predefined area of the browser application interface or network resource displayed in the browser application

[0089] In some embodiments of the present technology, the content discovery module 115 may be configured to perform a "crawler" operation. In other words, the content discovery module 115 can perform the work of a robot that “visits” multiple resources (eg, first network resource 124, second network resource 126, and multiple additional network resources 128) and catalogs one or more digital items located on the corresponding one of a first network resource 124, a second network resource 126, and a plurality of additional network resources 128. In some embodiments of the present technology, the content discovery module 115 may catalog digital items into an inverted index, matching the digital item to compile a list of keywords associated with the digital item ...

[0090] As part of the crawler function, the recommended content selection module 116 is configured to contain (on the storage device 118) information representing new found network resources and / or new found available content available thereon. In some embodiments of the present technology, the recommended content selection module 116 may be configured to maintain an inverted index on the storage device 118, but the recommended content selection module 116 may organize information representing new found network resources and / or new found content available on them into a data structure other than the inverted index.

[0091] In other embodiments of the present technology, instead of properly executing content discovery module 115, recommendation server 112 may share the functions of content discovery module 115 with another server (not shown) and / or another service (not shown). For example, the functions of the content discovery module 115 may be shared with a search engine server (not shown) that performs a search engine service. When the content discovery module 115 crawls and indexes new resources that may potentially contain text or other digital elements, the content discovery module 115 can also index such newly discovered (or updated) digital elements for the purposes of the recommendation server 112 procedures as described herein.

[0092] Recommended content selection module 116 may be configured to execute one or more machine learning algorithms (MLA). In some embodiments of the present technology, one or more machine learning algorithms may be any suitable supervised learning or semi-supervised learning algorithm, such as, for example:

• Artificial neural network

• Bayesian statistics

• Gaussian regression process

• Decision trees

• Etc

[0093] In general, the recommended content selection module 116 performs one or more MLAs to parse the indexed content items (i.e., those that are discovered and indexed by the content discovery module 115) to select one or more content items as the recommended content for user 102.

[0094] It should be noted that although the content detection module 115 and the recommended content selection module 116 have been described as separate elements, each of which performs corresponding functions, in other embodiments of the present technology, the functions performed by the content detection module 115 and the recommended content selection module 116 may be performed by a single element (eg, processing module 114). Alternatively, the functions performed by the content detection module 115 and the recommended content selection module 116 may be distributed among more modules than those shown in FIG. 1, and may run as part of multiple copies of the recommendation server 112.

[0095] In addition, each of the content detection module 115 and the recommended content selection module 116 may perform additional functions (ie, different from the respective functions described herein).

[0096] Continuing with FIG. 1 and referring to FIG. 3, in accordance with embodiments of the present technology, the featured content selection module 116 is configured to receive an indication of a plurality of user-item interactions 140, each user-item interaction of the plurality of user-item interactions 140 associated with a user 142 and a digital item 144.

[0097] In some embodiments of the present technology, the recommended content selection module 116 is configured to receive an indication of the plurality of user-item interactions 140 from the storage device 118. In some embodiments of the present technology, an indication of the plurality of user-item interactions 140 may be tracked and stored on storage device 118 by content discovery module 115. For example, the content discovery module 115 may be configured to download logs from a search engine server (not shown), logs from a browser server (not shown), and so on. Broadly, an indication of the plurality of user-to-item interactions 140 may be based on user browsing logs, user search logs, a combination of both, and so on.

[0098] In alternative embodiments of the present technology, the featured content selection module 116 is configured to obtain an indication of a plurality of user-to-item interactions 140 from an external source (not shown), such as a search engine log (not shown), a web browser log ( not shown) and so on.

[0099] In some embodiments of the present technology, digital item 144 may be one of a plurality of digital items 163. In general, any single digital item 144 of a plurality of digital items 163 may include any digital content (e.g., video, audio , text, their combination, etc.). In a specific non-limiting embodiment of the technology, any single digital item 144 of the plurality of digital items 163 may include textual content such as articles, books, other texts, and so on.

[00100] In some embodiments of the present technology, user 142 may be one of a plurality of users 148 (including user 102).

[00101] This indication of the plurality of user-item interactions 140 may include one or more of: time spent interacting with digital item 144, downloading digital item 144, reposting digital item 144, bookmarking digital item 144, leaving a comment associated with digital item 144, like digital item 144, and update digital item 144.

[00102] It should be noted that a given digital item 144 of a plurality of digital items can interact with a number of users 142 of a plurality of users 148. Similarly, a given user 142 of a plurality of users 148 can interact with more than one digital item 144 of a plurality of digital items 163 Naturally, a given user 142 out of a plurality of users 148 interact with one digital item 144 of a plurality of digital items 163 more than once.

[00103] It may also be that a given user 142 of the plurality of users 148 may not interact with all of the digital items 144 of the plurality of digital items 163. Similarly, it may be that a given digital item 144 of the plurality of digital items may not interact with all 142 users out of 148 users.

[00104] Recommended content selection module 116 is further configured, based on a plurality of user-item interactions 140, to determine how relevant a given digital item 144 is to a given user 142. More specifically, recommended content selection module 116 is configured based on a plurality of interactions 140 user-item, to create a matrix of user-item relevance scores.

[00105] With reference to FIG. 4 is a schematic diagram of a user-item relevance rating matrix 400. The presented user-item relevance score matrix 400 includes a plurality of columns 404 and a plurality of rows 402. Each of the plurality of columns 404 is associated with a given numeric item 144 of a plurality of numeric items 163. Each row of a plurality of rows 402 is associated with a given user 142 of the set users 148.

[00106] The overlap between a given row of a plurality of rows 402 and a given column of a plurality of columns 404 stores the relevance score user-item of the corresponding given user 142 out of the plurality of users 148 and the corresponding given digital item 144 of the plurality of digital items 163. For example, a particular row 408 is related with user "User 3", a particular column 406 is associated with the digital item "Item 3", and a particular cell 410 contains a relevance score of "5" indicating user 3 interactions with the digital item "Item 3", which indicates a rating relevance "5". How the relevance score is generated by the featured content selection module 116 is not particularly limited. For example, the featured content selection module 116 may generate a user-to-item relevance score using explicit user ratings (e.g., a given digital item 144 may be rated by users 142 on a predetermined scale, such as 1-5, 1-10, A-C, and etc). Alternatively, a user-to-item relevance score can be calculated based on a weighted algorithm. Alternatively, user-to-item relevance scores can be determined by a machine learning algorithm (not shown) by analyzing one or more measurements of user interaction (eg, latency, number of shares, and so on).

[00107] This cell 412 and this cell 414 do not contain any values. Thus, it can be argued that the user "User 1" did not interact with the digital item "Item 4" (as a result, this cell 412 is empty) and that the user "User 3" did not interact with the digital item "Item 6" (as a result, this cell 414 is blank). Thus, in the illustrated FIG. 4 depicts how the user-item relevance score matrix 400 can be "sparse". To avoid any doubt of interpretation, the term "sparse" describes the fact that some cells (ie, given cell 422, given cell 414 among others) are empty (ie, not containing a user-to-item relevance score). Empty cells are a sign of a lack of user-item interactions between the corresponding user and the corresponding digital item.

[00108] Recommended content selection module 116 is further configured to decompose the user-item relevance rating matrix 400 into a user-item matrix and an item matrix. FIG. 5 is a schematic representation of the decomposition of a user-element relevance score matrix 400 into a user-element matrix 502 and an element matrix 504. In some embodiments of the present technology, the recommended content selection module 116 may execute the SVD algorithm to decompose the user-to-member relevance score matrix 400 into a user-to-member matrix 502 and an element matrix 504. More specifically, a singular value decomposition (SVD) algorithm may use the aforementioned Formula 1 to split the user-item relevance score matrix 400 into a user-item matrix 502 and an item-matrix 504.

[00109] In some embodiments of the present technology, to perform the decomposition analysis, the recommended content selection module 116 may execute the ALS algorithm to decompose the user-item relevance score matrix 400. In some embodiments of the present technology, the ALS algorithm uses Formula 2 to represent a matrix 400 of user-item relevance scores.

[00110] In accordance with embodiments of the ALS algorithm methods using Formula 2, an initial version of the user matrix 502 (denoted as "U") and an initial version of the matrix 504 of elements (denoted as "V") are created (i.e., "initialization" performed by the recommended content selection module 116).

[00111] The user matrix 502 (denoted as "U") may be initialized in accordance with known techniques. For example, the recommended content selection module 116 may initialize the user matrix 502 with user vector values that are either randomly populated or filled with zeros (or a combination of both). In other words, the recommended content selection module 116 may generate initial user vector values for the user matrix 502 as random numbers or zeros (or a combination of both).

[00112] In accordance with embodiments of the present technology, the featured content selection module 116 is further configured to initialize the element matrix 504 with element vectors, the element vectors have been created such that digital elements with similar content have similar element vectors. It will be appreciated that in some embodiments of the present technology, numeric element 144 may be a text numeric element 144. In these embodiments of the present technology, the featured content selection module 116 is further configured to initialize the element matrix 504 with element vectors, the element vectors were are designed in such a way that elements with similar content have similar element vectors. In some embodiments of the present technology, the featured content selection module 116 is configured to create element vectors using word embedding techniques. Without limitation, the word embedding technique used may be at least one of word2vec or latent Dirichlet placement (LDA) techniques.

[00113] Recommended content selection module 116 is further configured to iteratively optimize an initial version of the user matrix 502 and an initial version of the element matrix 504. To this end, the recommended content selection module 116 is configured to iteratively: (i) first fix the user matrix 502 and adjust the element matrix 504, and (ii) further fix the element matrix 504 and customize the user matrix 502 ("one iteration"). After each iteration, the decomposition of the user matrix 502 and the element matrix 504 becomes closer to the user-to-element relevance score matrix 400. In some embodiments of the present technology, the recommended content selection module 116 is configured to iteratively optimize until a predetermined value of a predetermined metric is obtained. In a particular embodiment of the technology, the predetermined metric is the standard deviation (RJVISE). The predetermined RMSE value that triggers the completion of the iterative optimization execution can be empirically determined. In alternative embodiments of the present technology, the predetermined metric is the mean absolute error (MAR). The predetermined MAR value that triggers the completion of the iterative optimization execution can be empirically determined.

[00114] Upon completion of the optimization of the element array 504 (to obtain the optimized element array 504), the recommended content selection module 116 is configured to store the optimized element array 504 in the storage device 118. Recommended content selection module 116 may be further configured to omit user matrix 502.

[00115] In accordance with embodiments of the present technology, the thus stored optimized element matrix 504 can be used by the recommended content selector 116 to select the recommended content for the user 102.

[00116] More specifically, upon receipt of a request for content recommendation 150 from the electronic device 104, the recommended content selection module 116 is configured to obtain a user profile associated with the electronic device 104. In a particular embodiment of the technology, the recommended content selection module 116 obtains a user profile profile associated with electronic device 104 from storage device 118. In other embodiments of the present technology, the featured content selection module 116 is configured to obtain a user profile associated with the electronic device 104 from another source (not shown), for example, by targeting a profile database, and so on.

[00117] In some embodiments of the method, the user profile is a vector based on the user's browsing history. Exactly how the featured content selector 116 creates the vector based on the user's browsing history is not particularly limited. By way of example, a similar technique is described in a pending Russian patent application entitled "METHOD AND SYSTEM FOR PROCESSING USER-RELATED ACTIVITY INDICATIONS" under No. 2015111633, filed on April 31, 2015; the content of the application is incorporated in its entirety into this application by reference.

[00118] Next, based on the user profile, the recommended content selection module 116 selects at least one recommended content item. It should be noted that the recommended content selection module 116 selects the recommended content item using the optimized item matrix 504 that is stored in the storage device 118.

[00119] More specifically, the recommended content selection module 116 is configured to reconstruct the user-to-item relevance rating matrix 400 using the optimized item-matrix 504 and the user profile to create the reconstructed user-item relevance rating matrix 400. In accordance with embodiments of the present technology, the recommended content selection module 116 reconstructs the reconstructed user-item relevance rating matrix 400, so that each user-item pair from the user-item relevance rating matrix 400 is associated with a corresponding user-item relevance rating, corresponding rating user-item relevance is the relevance of a given digital item 144 to a given user 142 (in this case, user 102).

[00120] Exactly how the suggested content selection module 116 selects a particular digital item 114 for the user 102 is not particularly limited. An example of a machine learning algorithm that can be used by the content recommendation selection module 116 is described in a pending Russian patent application entitled METHOD AND APPARATUS FOR GENERATING A RECOMMENDED CONTENT LIST "METHOD AND APPARATUS FOR GENERATING A RECOMMENDED CONTENT LIST" 2015136684, filed August 22, 2015; the contents of which are incorporated herein in full by reference.

[00121] Naturally, the recommended content selection module 116 can select more than one digital item 114 for the user 102. For example, the recommended content selection module 116 can select a plurality of digital items 144 for use as a subset of the recommended digital items 144 for the user 102. After a content recommendation selection module 116 creates a subset of content items for a given user 102, a content recommendation selection module 116 creates a content recommendation message 152 and transmits it to the electronic device 104.

[00122] When the electronic device 104 receives the content recommendation message 152, the recommendation application 106 processes its content and displays the content recommendation items on the recommendation interface 108. How the recommendation interface 108 is implemented is not particularly limited - the recommendation interface 108 shown in FIG. 2. In another example, the recommendation interface 108 can be implemented as described in a pending Russian patent application entitled: "METHOD AND SYSTEM FOR INTERACTING WITH A CONTENT ELEMENT" (ENGLISH "METNO D OF AND SYSTEM FOR INTERACTING WITH A CONTENT ELEMENT A CONTENT STREAM ") No. 2015141291, filed September 29, 2015; the contents of which are incorporated herein in full by reference.

[00123] Given the described architecture and the examples provided, it is possible to perform a machine learning algorithm (MLA) training method to select content recommendations for an electronic device user.

[00124] FIG. 6 is a flow diagram of a method 600 implemented in accordance with non-limiting embodiments of the present technology. The method 600 can be performed on the recommendation server 112, and more specifically, by the recommended content selection module 116.

[00125] 602 - Obtaining an indication of a plurality of user-item interactions, each user-item interaction is associated with a user and a digital item

[00126] The method 600 begins at block 602, where the recommended content selector 116 is instructed to a plurality of user-item interactions 140, each user-item interaction 140 associated with a user 142 and a digital item 144.

[00127] In some embodiments of the method 600, the step of obtaining an indication of the plurality of user-item interactions 140 includes obtaining an indication of the plurality of user-item interactions from the user's browsing histories. A user-item interaction is: time spent interacting with a digital item, uploading a digital item, reposting a digital item, bookmarking a digital item, leaving a comment associated with a digital item, liking a digital item, and updating a digital item.

[00128] 604 - based on a plurality of user-item interactions, create a matrix of user-item relevance scores

[00129] At block 604, the recommended content selection module 116, based on the plurality of user-item interactions 140, creates a matrix 400 of user-item relevance scores.

[00130] 606 - decomposition of a matrix of user-element relevance scores into a matrix of users and matrices of elements, the decomposition includes: initializing the matrix of elements using vectors of elements, the vectors of elements were created in such a way that digital elements with similar content have similar vectors of elements, initialization of the user matrix using user vectors.

[00131] In step 606, the recommended content selection module 116 decomposes the user-item relevance score matrix 400 into a user-item matrix 502 and an item matrix 504. In accordance with embodiments of the present technology, the decomposition step includes: initializing the element matrix 504 with element vectors, the element vectors have been created such that digital elements with similar content have similar element vectors. Recommended content selection module 116 further initializes user matrix 502 with user vectors.

[00132] In some embodiments of the method 600, initializing the user matrix 502 with user vectors includes populating the user matrix 502 with random initial values of the user vectors.

[00133] In some embodiments of method 600, the step of initializing a user matrix with user vectors includes populating the user matrix with zero user vector values.

[00134] In some embodiments of the method 600, the numeric item is a text numeric item and the item vectors were created using the word embedding method. The step of creating element vectors may be performed with a word embedding method, the word embedding method is at least one of word2vec or latent Dirichlet placement (LDA).

[00135] In some embodiments of the method 600, the decomposition step is performed using a singular value decomposition (SVD) analysis. Decomposition further includes decomposition analysis, decomposition analysis is performed using an alternating least squares (ALS) algorithm.

[00136] 608 - Iterative optimization of a user matrix and an element matrix.

[00137] In block 608, the recommended content selection module 116 iteratively optimizes the user matrix 502 and the element matrix 504.

[00138] In some embodiments of the method 600, the step of iteratively optimizing the user matrix 502 and the element matrix 503 is performed until a predetermined value of a predetermined value is reached. The predefined metric can represent (without setting limits): Root mean square deviation (RMSE) or mean absolute error (MAR).

[00139] 610 - saving an optimized matrix of elements

[00140] At block 610, the recommended content selection module 116 stores the optimized element matrix 504.

[00141] In some embodiments of the method 600, after iterative optimization, the user matrix 502 is omitted.

[00142] Using the optimized element matrix 504 to select the recommended digital element 144

[00143] As previously mentioned, after the recommended content selection module 116 stores the optimized item matrix 504, the optimized item matrix 504 can be used to create the recommended digital item 144 for the user 102. To do this, the method 600 further includes: receiving from the electronic device 104 a request for recommended content 150, obtaining a user profile associated with the electronic device 104; and selecting at least one recommended content item 144 based on the user profile and the optimized item matrix 504.

[00144] In some embodiments of the method, the selection step includes reconstructing the user-to-member relevance matrix 400 with the optimized member matrix 504 and the user profile to create the reconstructed user-member relevance matrix 400.

[00145] In some embodiments of the method, each user-item pair of the reconstructed user-item relevance score matrix 400 is associated with a corresponding user-item relevance score, the corresponding user-item relevance score represents the relevance of the given numeric item 144 to user 102.

[00146] In some embodiments of the method, the user profile is a vector based on the user's browsing history.

[00147] It is important to keep in mind that not all of the technical results mentioned herein may be manifested in every embodiment of the present technology. For example, embodiments of the present technology can be performed without exhibiting some of the technical results, others can be executed with or without other technical effects.

[00148] Modifications and improvements to the above-described embodiments of the present technology will be apparent to those skilled in the art. The foregoing description is provided by way of example only and does not establish any limitation. Thus, the scope of the present technology is limited only by the scope of the appended claims.

[00149] Embodiments of the present technology may be summarized in numbered paragraphs.

[00150] ITEM 1 Method (600) for teaching a machine learning algorithm to select recommended content for a user (102) of an electronic device (104), the method is performed on a recommendation server (112) accessible to an electronic device via a data transmission network (SW), the recommendation server executes post-learning machine learning algorithm; the method includes:

[00151] obtaining (602) an indication of a plurality of user-item interactions (140), each user-item interaction associated with a user (142) and a digital item (144);

[00152] based on the plurality of user-item interactions, creating (604) a matrix (400) of user-item relevance scores;

[00153] decomposition (606) of the user-element relevance assessment matrix into a matrix (502) of users and a matrix (504) of elements, the decomposition includes:

[00154] initializing a matrix of elements using vectors of elements, the vectors of elements were created in such a way that digital elements with similar contents have similar vectors of elements,

[00155] initializing the user matrix with user vectors;

[00156] iterative optimization (608) of the user matrix and the element matrix;

[00157] storing (610) the optimized matrix of elements.

[00158] ITEM 2 The method of claim 1, further comprising:

[00159] upon receipt of a request for recommended content from an electronic device (150),

[00160] obtaining a user profile associated with the electronic device; and

[00161] selection of at least one recommended content item (144), selection is made based on user profile and optimized matrix of items.

[00162] ITEM 3 The method of claim 2, wherein the selection includes reconstructing a user-member matrix of ratings with an optimized member matrix and a user profile to create a reconstructed user-member matrix.

[00163] ITEM 4 The method of claim 3, wherein each user-item pair of the reconstructed user-item matrix is associated with a corresponding user-item relevance score, the corresponding user-item relevance score represents the digital item's relevance to the user.

[00164] ITEM 5 The method of claim 2, wherein the user profile is a vector based on the user's browsing history.

[00165] ITEM 6 The method according to any one of claims. 1-5, in which initializing the user matrix with user vectors includes populating the user matrix with random values of the user vectors.

[00166] ITEM 7 The method according to any one of claims. 1-5, in which initializing the user matrix with user vectors includes populating the user matrix with the initial values of the user vectors being zero.

[00167] ITEM 8 The method according to any one of claims. 1-7, in which, after said iterative optimization, the user matrix is omitted.

[00168] ITEM 9 The method according to any one of claims. 1-8, in which the numeric item is a text numeric item, and wherein the item vectors have been generated using the word embedding method.

[00169] CLAUSE 10 The method of claim 9, which further includes generating vectors of elements with a word embedding method, the word embedding method is at least one of word2vec or Latent Dirichlet Allocation (LDA).

[00170] ITEM 11 The method according to any one of claims. 1-10, in which the decomposition is performed using a singular value decomposition (SVD) analysis.

[00171] ITEM 12 The method of claim 11, wherein the decomposition further includes decomposition analysis that is performed using an alternating least squares (ALS) algorithm.

[00172] ITEM 13 The method according to any one of claims. 1-12, wherein said obtaining an indication of a plurality of user-item interactions includes obtaining an indication of a plurality of user-item interactions from user browsing histories.

[00173] ITEM 14 The method of claim 13, wherein the user-item interaction is: time spent interacting with the digital item, downloading the digital item, reposting the digital item, bookmarking the digital item, leaving a comment associated with the digital item , leaving a “like” to the digital item and updating the digital item.

[00174] ITEM 15 The method according to any one of paragraphs. 1-14, in which iterative optimization of the user matrix and the element matrix is performed until a predetermined value of a predetermined value is reached.

[00175] ITEM 16 The method of claim 15, wherein the predetermined metric is one of the following: Standard deviation (RMSE) and mean absolute error (MAR).

[00176] ITEM 17 Server (112), including:

[00177] a storage medium (118);

[00178] a network interface capable of being connected over a data network;

[00179] a processor (114), functionally connected to a storage medium and a network interface, the processor being configured to perform a method according to any one of claims. 1-16.

13. The method of claim 1, wherein receiving an indication of the plurality of user-item interactions includes obtaining an indication of the plurality of user-item interactions from user browsing histories.

14. The method according to claim 13, wherein the user-item interaction is: time spent interacting with the digital item, downloading the digital item, reposting the digital item, adding the digital item to bookmarks, leaving a comment associated with the digital item, leaving like "digital item and updating the digital item.

15. The method of claim 1, wherein the iterative optimization of the user matrix and the element matrix is performed until a predetermined value of the predetermined value is reached.

16. The method of claim 15, wherein the predetermined value of the predetermined magnitude is one of the following: Standard deviation (RMSE) and mean absolute error (MAR).

17. The server includes:

data carrier;

a network interface configured to connect over a data network;

a processor functionally connected to a data carrier and a network interface, the processor is configured to:

obtaining an indication of a plurality of user-item interactions, each user-item interaction associated with a user and a digital item; based on a variety of user-element interactions, creating a matrix of user-element relevance scores;

decomposition of the user-element relevance assessment matrix into a matrix of users and a matrix of elements, the decomposition includes:

initialization of the matrix of elements using vectors of elements, the vectors of elements were created in such a way that digital elements with similar content had similar vectors of elements, initialization of the matrix of users using vectors of users;

iterative optimization of the user matrix and the element matrix,

storing the optimized matrix of elements in the data carrier.

Claims (33)

1. A method for training a machine learning algorithm, running on a recommendation server accessible to an electronic device via a data transmission network, the recommendation server executes the machine learning algorithm after training; a way that includes: obtaining an indication of a plurality of user-item interactions, each user-item interaction associated with a user and a digital item; based on a set of user-element interactions, creating a matrix of user-element relevance scores; decomposition of the matrix of user-element relevance assessments into a matrix of users and a matrix of elements, the decomposition includes: initialization of the matrix of elements using vectors of elements, the vectors of elements were created in such a way that digital elements with similar content had similar vectors of elements, initialization of the matrix of users using vectors of users; iterative optimization of the user matrix and the matrix of elements; saving an optimized matrix of elements. 2. The method according to claim 1, further comprising: upon receipt of a request for recommended content from an electronic device, obtaining a user profile associated with the electronic device; and selection of at least one recommended content item, selection is made based on user profile and optimized item matrix. 3. The method of claim 2, wherein the selection includes reconstructing the user-member rating matrix with the optimized member matrix and the user profile to create the reconstructed user-member matrix. 4. The method of claim 3, wherein each user-item pair of the reconstructed user-item matrix is associated with a corresponding user-item relevance score, the corresponding user-item relevance score represents the digital item's relevance to the user. 5. The method of claim 2, wherein the user profile is a vector based on the user's browsing history. 6. The method of claim 1, wherein initializing the user matrix with user vectors includes populating the user matrix with random values of the user vectors. 7. The method of claim 1, wherein initializing the user matrix with the user vectors includes populating the user matrix with initial values of the user vectors equal to zero. 8. The method of claim 1, wherein after said iterative optimization, the user matrix is omitted. 9. The method of claim 1, wherein the digital item is a text numeric item and wherein the item vectors have been generated using the word embedding method. 10. The method of claim 9, further comprising creating vectors of elements using a word embedding method, the word embedding method is at least one of word2vec or Latent Dirichlet Allocation (LDA). 11. The method of claim 1, wherein the decomposition is performed using a singular value decomposition (SVD) analysis. 12. The method of claim 11, wherein the decomposition further includes decomposition analysis that is performed using an alternating least squares (ALS) algorithm. 13. The method of claim 1, wherein receiving an indication of the plurality of user-item interactions includes obtaining an indication of the plurality of user-item interactions from user browsing histories. 14. The method according to claim 13, wherein the user-item interaction is: time spent interacting with the digital item, downloading the digital item, reposting the digital item, adding the digital item to bookmarks, leaving a comment associated with the digital item, leaving like "digital item and updating the digital item. 15. The method of claim 1, wherein the iterative optimization of the user matrix and the element matrix is performed until a predetermined value of the predetermined value is reached. 16. The method of claim 15, wherein the predetermined value of the predetermined magnitude is one of the following: standard deviation (RMSE) and mean absolute error (MAR). 17. The server includes: data carrier; a network interface configured to connect over a data network; a processor functionally connected to a data carrier and a network interface, the processor is configured to perform: obtaining an indication of a plurality of user-item interactions, each user-item interaction associated with a user and a digital item; based on a set of user-element interactions, creating a matrix of user-element relevance scores; decomposition of the matrix of user-element relevance assessments into a matrix of users and a matrix of elements, the decomposition includes: initialization of the matrix of elements using vectors of elements, the vectors of elements were created in such a way that digital elements with similar content had similar vectors of elements, initialization of the matrix of users using vectors of users; iterative optimization of the user matrix and the matrix of elements; storing the optimized matrix of elements in the data carrier.

Images